Artificial General Intelligence and Noncomputability: A Dynamical Framework

Abstract

Achieving genuine (human-level) artificial general intelligence (AGI) is one of the major goals of computer science, engineering, psychology, neuroscience, and mathematics. In this paper, we critically reexamine the relation between natural intelligence and artificial intelligence at a fairly general theoretical level. After identifying four major structural themes in natural intelligence, we move to the issue of AGI implementation through physical computing machines. Motivated by Penrose’s Gödelian argument refuting the thesis of AGI realizability via Turing machines, we formulate several theses on the noncomputable character of AGI systems. In particular, we support the claim that infinitary noncomputability might constitute a viable path toward future AGI implementations, especially if coupled with nonlocality and a nonclassical probabilistic structure such as those in the quantum world. A theoretical mathematical framework for realizing AGI through non-Markovian stochastic dynamic systems is then presented and illustrated by describing multi-agent AGI assemblages comprised of interconnected dynamic agents. We envision that such networked dynamical assemblages might be powered by noncomputable physics or arranged in an infinitary structure.